Accurate extraction of roads from remote sensing images has been a hot research topic, but challenges remain due to complex spatio-temporal scenes and multi-scale road objects. In this paper, we propose MC-DeepLab, a novel road extraction network based on Mamba. Specifically, we introduce VMamba (VM) block to extend DeepLabv3+. The VM block integrates a cross-scanning module that expands images into sequences from four directions and inputs them into Mamba module for generating global contextual information. To encode Multi-scale feature representation, we construct a Mamba spatial pyramid pooling (Mamba-SPP). This design leverages Mamba’s powerful sequence modeling capabilities while capturing local image dependencies through split windows. We also introduce coordinate attention to reduce redundant information in multi-scale features to improve the accuracy of road extraction. Comprehensive experiments on two public datasets show that the proposed method outperforms the state-of-the-art methods, demonstrating the effectiveness and potential of MC-DeepLab.